The elusive ROI of AI in healthcare

AI in healthcare holds great promise, but promises don’t pay the bills. In Bain And McKinsey Surveys show that US healthcare leaders expect – and even need – a positive return on their AI investments. However, demonstrating ROI is a challenge and will remain a major barrier to adoption for the foreseeable future.

The potential benefits and costs of AI

Healthcare organizations could use AI for a variety of benefits, including improving the quality of care, improving patient and staff experiences, accelerating research and gaining more insights from data.

They can also use AI to directly increase revenue by increasing volume, accelerating throughput, increasing risk adjustment and service-level coding, and improving revenue cycle management. Meanwhile, AI could reduce costs by reducing workforce needs, reducing employee turnover and improving supply chain efficiency.

Of course, AI also entails costs. Evaluating different AI products takes time and effort, and organizations must consider opportunity costs (AI can distract from other activities) and reputational risks (due to potential adverse events).

Moreover, implementing AI is complexrequires significant resources and is fraught with potential pitfalls. Productivity often drops temporarily (“switching costs”). Once implemented, there will be ongoing expenditures on software, monitoring and data infrastructure.

Why analyzing AI ROI is so difficult

Translating abstract concepts such as quality, efficiency and productivity into numbers is a challenge and requires asking many difficult questions.

First, from whose perspective is ROI assessed? The interests of the various stakeholders in healthcare are not always aligned. For example, a nurse manager may find AI that triages patients attractive if it reduces staffing needs, but patients may be wary of having to do more work alone.

Second, who pays for AI, and who influences it? Often, those who make AI purchasing decisions differ from those who have the most influence over them. For example, a top-level executive may push an AI tool that increases risk coding, but physicians may resist if they need to change the way they document care.

Third, what is the time horizon? History tells us that organizations primarily use technology to make their existing processes more efficient, limiting the overall benefits. It takes many years to find new, better ways to produce goods and services. We see this today as healthcare systems rush to adopt AI that allows doctors to write the same (usually sloppy) notes more quickly rather than completely rethinking clinical documentation.

Fourth, what is the basic performance? Most organizations know little about the time and effort that physicians and staff spend on various tasks (e.g., writing a discharge summary) and even less about the quality of their results (e.g., are the summaries accurate and readable?).

Finally, which metrics best evaluate the impact of AI? Healthcare data is siled and incomplete, and it is difficult to measure and assign value to constructs such as quality and physician well-being.

Why a positive ROI can be elusive

Jim Covello, Head of Global Equity Research at Goldman Sach, explained: “The significant costs of developing and operating AI technology mean that AI applications must solve extremely complex and important problems for companies to achieve appropriate returns on their investments.” But it may be too much to expect that today’s AI can solve complex and important healthcare problems.

First, generative AI tools are typically too unreliable and error-prone to apply to high-end tasks. So most organizations use them to take the burden off of ‘boring’ tasks such as writing clinical notes and completing prior authorizations. Yet AI could paradoxically make these tasks more difficult. Doctors at UC San Diego Health who used ChatGPT to respond to patient messages paradoxically passed them on 22% more time on this task than those who did not use AI.

Even AI solutions that save time may not increase productivity. In keeping with Parkinson’s law – that ‘work expands to fill the time allotted for its completion’ –three out of four British doctors reported that this would be the case not spend time that AI frees up on caring for patients. American doctors now adopting tools like AI writers are likely to be no different.

Likewise, AI does not work in a vacuum. Organizations must remove several downstream limitations to realize the benefits of AI. For example, automating patient scheduling will not improve access if physicians already have full schedules. Likewise, algorithms that identify inpatients ready for discharge are useless if there is no way to send patients after hospital admission.

Payment models bring additional challenges. Most healthcare payments are fee-for-servicewith payers occurring very rarely reimburse AI software. To break even financially, organizations that adopt AI must increase service volumes, often substantially, given their low single-digit operating margins. Yet most AI tools — for example, those that summarize clinical records, identify patients at high risk for deterioration, or help detect precancerous colon polyps — do not impact volume.

Across all industries, technology transformation programs typically achieve less than a third of their expected value. AI in healthcare will be no different.

Looking ahead

None of this is to say that AI in healthcare is worthless. AI can make healthcare more accessible, effective and sustainable. Still, ROI pressure will increase as AI hype (and the associated FOMO) subsides.

Consequently, AI will penetrate most quickly into back-office financial areas, such as revenue cycle management. Startups this is an area where we have already seen some of the highest maturities, valuations and exits.

Likewise, many physician-focused AI products will extend into activities that directly impact finances. For example, documentation and summary tools will start recommending risk codes and costs.

For other AI products, organizations may require clinical teams to see more patients or reduce staffing levels. The point is that while AI holds promise and can transform healthcare in the long term, there will be no free lunch.